A Hybrid Symbolic-Connectionist Approach to Modeling Emotions Randolph M. Jones Eric Chown

From: AAAI Technical Report FS-01-02. Compilation copyright © 2001, AAAI (www.aaai.org). All rights reserved.
A Hybrid Symbolic-Connectionist Approach to Modeling Emotions
Randolph M. Jones
Eric Chown
Amy E. Henninger
Senior Scientist & Assistant Professor
Soar Technology & Colby College
5847 Mayflower Hill Drive
Waterville, ME 04901-8858
(207)872-3831
rjones@soartech.com, rjones@colby.edu
Assistant Professor of Computer Science
Bowdoin College
8650 College Station
Brunswick ME 04011
207-725-3084
echown@bowdoin.edu
Senior Scientist
Soar Technology
317 N. First St.
Ann Arbor, MI
321 276 5135
amy@soartech.com
Introduction
This paper describes a framework for modeling emotions
in an interactive, decision-making agent. In tune with
modern theories of emotions (e.g., Damasio, 1995;
LeDoux, 1992), we regard emotions essentially as
subconscious signals and evaluations that inform, modify,
and receive feedback from a variety of sources including
higher cognitive processes and the sensorimotor system.
Thus, our work explicitly distinguishes the subconscious
processes (in a connectionist implementation) and the
decision making that is subject to emotional influences (in
ModSAF
Soar
Long-term memory
Decision making
Emotional Appraisal
Working memory
Soar-ModSAF
Interface
Clarity
Pleasure
Confusion
Pain
Arousal
Intensity
Soar Emotions Interface
Output Commands
Perceptual Analysis
Environment
Behaviors
RWA Command Agent
PDUs
HLA Gateway
Figure 1. The hybrid architecture for emotional behavior.
a symbolic cognitive architecture). Because our project
focuses upon decision making, it emphasizes aspects of
emotion that influence higher cognition and not those that
affect, for example, the immune system.
We are
integrating a connectionist model of emotions from Chown
(1993) with Rosenbloom, Laird, and Newell’s (1993;
Newell, 1990) Soar architecture. Sponsored by the Army,
the application area incorporates emotions and individual
differences into behavior models of synthetic virtual pilots
in a battlefield simulation. Intelligent agents in this
application area must exercise a variety of reasoning
capabilities, including situation assessment, planning,
reacting to goal failures, and interacting with a team of
agents. Although we are developing the framework for this
model within the military domain, we intend the
framework to generalize across interactive agents. Figure
1 provides a sketch of the integrated architecture.
In our framework, symbolic assessments of a small set
of “emotional attributes” reside in a working memory,
which serves as the interface between deliberative
cognitive processes and the emotion mechanisms. Working
memory elements combine with background knowledge to
generate strategies, reasoning, and external behavior, as
well as working interpretations of the environment and
status of internal goals (situational awareness). Some of
these interpretations and assessments feed into the
connectionist model, which in turn continuously computes
new values for each emotional attribute. This paper
presents a work in progress.
We have begun
implementation of the architecture on top of an existing
Soar model, but have not yet begun testing.
The Connectionist Component
In the current framework, the connectionist model consists
of several interacting components, including an arousal
level system, a pleasure/pain system (including both
physical and cognitive inputs), and clarity and confusion
mechanisms. These are the emotional attributes that
combine to influence cognition in a variety of ways. We
are intentionally avoiding any explicit notion or labels of
“emotional state”, assuming instead that such states are
post-hoc assessments of trends in behavior that arise from
various combinations of arousal, pleasure, pain, etc.
Different qualitative levels of attribute values generated by
the connectionist model ultimately influence how the
cognitive architecture reasons, makes decisions, perceives,
and acts. The pleasure/pain system interprets the level to
which a stimulus represents a threat or enhancement to
survival. This applies both to immediate sensations of
physical pain and to more deliberate predictions of
situations and outcomes. In turn, pleasure and pain
stimulate the arousal system. As indicated in Figure 2, the
pleasure/pain continuum is also dependent on the
individual’s sense of clarity versus confusion in situational
awareness. The figure is a simplification as there are
numerous other inputs to each of the units shown (e.g.
direct sensation of pain, or naturally arousing stimuli). Our
incorporation of the influences of cognitive confusion and
clarity are due to a model postulated by Kaplan (1991),
which we do not have space to elaborate here.
Arousal
Pain
Pleasure
We should stress that we are not developing a full-fledged
theory of personality. Rather, because emotion serves as
one important part of personality, we are simply ensuring
that our model of emotions is consistent with currently
accepted frameworks for personality.
Additional analysis will relate our framework to more
traditional, symbolic assessments of emotion (e.g., Ortony,
Clore, & Collins, 1998). Our intention is first to map our
emotional attributes to more familiar emotional state labels.
We can then vary parameters to explore behavior in the
system that could be characterized as, for example, “angry”
or “fearful”. Our evaluation will also explore the effects of
background knowledge on emotional reasoning. Because
arousal serves as a filter on retrieval and applicability of
long-term knowledge, we can certainly expect to observe
behavior differences based on long-term knowledge
differences. It will be a challenge to do a systematic study
along these lines, but we will attempt to test agents that
include “typical” knowledge differences that might arise
from differences in training and background experience.
Acknowledgements
Confusion
Clarity
(Low-Threshold Activity)
(High-Threshold Activity)
This work was sponsored by the U.S. Army Research
Institute’s Emotional Synthetics Forces STTR – Ph II,
contract number: DASW01-99-C-0037.
Figure 2. Sketch of computational arousal mechanism
References
The Symbolic Component
The decision making agent is based on interactive real-time
expert systems that are used for training simulations by the
US military (Jones et al., 1999; Hill et al., 1997). As a
scenario unfolds, a command agent reactively plans and
communicates with its subordinates. Our work will
parameterize the commander to make it susceptible to the
emotional attributes from the connectionist subsystem. In
addition, the commander knowledge will incorporate an
appraisal system and a response system. The response
system accepts behavior moderators from the connectionist
model. As the agent monitors its progress (and the
progress of its teammates), the appraisal system signals
events that feed into the connectionist system.
Experimental Plan
The experimental plan for our system is still being
designed. One empirical investigation will examine our
model’s relationship to a variety of “personality types” that
vary along emotional lines. The connections in our
emotions subsystem are parameters affecting things such as
how sensitive an individual is to pain or arousal. Different
parameter values will result in different patterns of
behavior. These individual differences lead to distinctive
decision making profiles that can lead to crucial
performance differences in emotionally charged situations.
Chown, E. 1993. Consolidation and learning: A
connectionist model of human credit assignment. Doctoral
Dissertation, The University of Michigan.
Damasio, A. R. 1995. Descartes’ error: Emotion, reason,
and the human brain. New York: Avon.
Hill et al. 1997. Intelligent agents for the synthetic
battlefield:A company of rotary-wing aircraft. Proceedings
of the Ninth Conference on Innovative Applications of
Artificial Intelligence. Menlo Park, CA: AAAI Press.
Jones et al. 1999. Automated intelligent pilots for combat
flight simulation. AI Magazine, 20(1), 27-41.
Kaplan, S. 1991.
Beyond rationality: Clarity-based
decision making. In Garling & Evans (Eds.), Environment,
Cognition, and Action: An Integrative Multidisciplinary
Approach. New York: Oxford University Press.
LeDoux, J. 1992. Brain mechanisms of emotion and
emotional learning. Current Opinions in Neurobiology, 2.
Milner, P.M. 1991. Brain stimulation reward: A review.
Canadian Journal of Psychology, 45, 1-36.
Newell, A. 1990. Unified theories of cognition.
Cambridge, MA: Harvard University Press.
Ortony, A., Clore, G. L., & Collins, A. 1988. The
cognitive structure of emotions. Cambridge, England:
Cambridge University Press.
Rosenbloom, P., Laird, J., and Newell, A 1993. The Soar
papers: Research on integrated intelligence. Cambridge,
MA: MIT Press.